基于表情分解-扭曲变形的人工表情合成算法

Facial Expression Synthesis Based on Facial Expression Decomposition and Warping

为了能快速有效地生成任意强度的人脸表情图像,提出了一种鲁棒的可以生成带任意强度表情图像的人工表情合成算法,该算法首先通过施加高阶奇异值分解(HOSVD)来把训练集分解为个人、表情和特征3个子空间,并把它们映射到表情子空间中,用来合成任意人脸正面照片的任意强度、任意表情的图像;在生成图像时,不采用通常所使用的线性组合基图像生成法,而是对源图像进行扭曲变形,这不仅能使训练数据和计算量大为减少,还可以生成任意尺寸、任意背景、任意光照、任意色彩或任意姿势的表情图像,且通过二次插值,还可以得到任意强度的表情图像。实验证明,该算法效率较高,且生成的图像效果很好。

This paper proposes a robust algorithm of facial expression synthesis. Decompose a train set into person subspace, expression subspaee and feature subspace by employing higher-order singular value decomposition （HOSVD）, and then map them to expression subspace, which can be used to synthesize an expressive image of arbitrary strength, given any frontal facial image. By warping source image to generate the target image rather than by linear-combining the images in train set as usual, this approach not only lessens the train set data storage and the computation complexity, but also enables the system to cope with facial images of arbitrary size, background, illumination condition, coloration or pose. With square interpolation, expression facial images of any assigned strength can be obtained. Experiments prove that this algorithm has higher efficiency and can generate very excellent images.